Graph Neural Networks: A Review of Methods and Applications

The paper is a review of graph neural networks (GNNs), connectionist models that capture the dependence within graph-structured data through message passing between nodes and, unlike standard neural networks, retain a state that can represent information from a node's neighborhood at arbitrary depth. It motivates GNNs with the many learning tasks that require graph inputs-such as modeling physical systems, learning molecular fingerprints, predicting protein interfaces, and classifying diseases-as well as reasoning over structures extracted from text and images, like dependency trees and scene graphs.

The survey notes that although primitive GNNs were difficult to train toward a fixed point, recent advances in network architectures, optimization techniques, and parallel computation have enabled successful learning, with variants such as graph convolutional networks (GCN), graph attention networks (GAT), and gated graph neural networks (GGNN) delivering ground-breaking performance across these tasks. To orient future work, the authors provide a detailed review of existing GNN models, systematically categorize their applications, and propose four open problems for further research.